Texas's AI law has been in force since January 1, 2026, the Attorney General's consumer complaint portal is live, and a violation that isn't cured within 60 days of notice can cost up to $200,000. If your compliance team built a risk-assessment program for Texas modeled on Colorado's original AI Act, you built the wrong program: TRAIGA's enacted text dropped that entire regime before it ever passed. What it kept is narrower, sharper, and easy to misjudge in either direction.
You're the person who has to tell your CEO whether Texas requires anything beyond what you already do for the EU AI Act or a Colorado-style framework. The honest answer is: almost nothing, unless your AI touches one of five specifically prohibited uses, or your organization is a government agency. Getting that distinction wrong in the cautious direction wastes budget on documentation nobody asked for; getting it wrong in the careless direction leaves a real prohibited-use exposure sitting on your books uninsured.
Key Takeaways
- TRAIGA (Texas House Bill 149, the Texas Responsible Artificial Intelligence Governance Act) took effect January 1, 2026, and prohibits five specific AI uses rather than imposing a general private-sector risk-assessment duty: the broad regime in earlier drafts was removed before enactment.
- The Texas Attorney General has exclusive enforcement authority, there is no private right of action, and violators get a mandatory 60-day cure notice before any penalty attaches; penalties then range from $10,000 to $200,000 per violation depending on curability, plus up to $40,000 per day for continuing violations.
- Most of TRAIGA's disclosure and transparency obligations bind government agencies, not private companies. The private-sector duty is narrower and behavior-based (the five prohibited uses), not a documentation-and-assessment regime like Colorado's original AI Act or the EU AI Act.
TRAIGA Is Not the Law Its Early Drafts Promised
The bill that became TRAIGA started life closer to a Texas version of the EU AI Act: earlier drafts included duty-of-care obligations, algorithmic discrimination liability, and impact-assessment requirements that would have applied broadly across the private sector. By the time HB 149 passed the Texas Legislature and Governor Abbott signed it, that structure was gone. What remains is an intent-based prohibition list plus a set of disclosure duties that fall mostly on state government.
Robert Brown, a Latham & Watkins partner who has advised clients on the law, put the practical effect plainly: "Governmental agencies will feel the greatest impact, as many of the requirements under the final version of the bill apply exclusively to them." That is the single fact most law-firm client alerts bury under paragraphs of legislative history, and it is the fact that determines whether your compliance program needs to change at all.
TRAIGA's sponsor, State Rep. Giovanni Capriglione, described the legislative intent behind the narrowing directly: "We didn't want to be overburdensome, and we wanted to try to do this in a way that is reasonable and protects people from some of the biggest harms." The result is a law that reads, in practice, more like a targeted misuse statute than a governance framework — closer to how Texas already regulates deceptive trade practices than to how the EU AI Act classifies and audits high-risk systems.
TRAIGA vs. Colorado vs. California: What Each State Actually Requires in 2026
| Requirement | Texas TRAIGA | Colorado AI Act (as revised) | California SB 942 |
|---|---|---|---|
| Effective date | January 1, 2026 | January 1, 2027 (delayed twice from an original February 2026 date) | August 2, 2026 (deferred by AB 853 to align with the EU AI Act's high-risk enforcement date) |
| Core structure | Prohibited-use list plus government disclosure duties | Originally a broad risk-based regime; the May 2026 rewrite (SB 189) dropped the duty of care, impact assessments, and risk-management-program mandate in favor of narrower disclosure/transparency rules | Transparency law for generative AI content: detection tools and watermarking, not a risk-classification regime |
| Who has the heaviest duties | Government agencies (consumer-facing AI disclosure) | Deployers of "high-risk" automated decision systems, though obligations are now lighter after the rewrite | Generative AI providers with more than 1 million monthly users |
| Private-sector risk assessments required? | No. The general risk-assessment regime was removed before enactment | Reduced substantially by the 2026 rewrite; no longer requires full impact assessments and risk-management programs as originally drafted | Not applicable: this is a content-transparency law, not a risk-assessment law |
| Enforcement | Texas AG, exclusive authority, no private right of action, mandatory 60-day cure notice | Colorado AG | California AG |
| Penalty range | $10,000–$200,000 per violation; $2,000–$40,000/day for continuing violations | Under revision as part of the 2026 rewrite | Civil penalties for missing detection tools or disclosure failures |
The pattern across all three: 2026 has been a year of state AI laws getting narrower, not broader, as they approach their effective dates. If you already track California's AI transparency obligations separately, treat TRAIGA as a distinct, smaller add rather than folding it into the same disclosure workflow. If your multi-state AI governance program was built to the ambition of 2024's draft bills rather than the 2026 enacted text, it is over-built for Texas and probably over-built for Colorado too.
The Five Prohibited AI Practices
TRAIGA's substantive core sits in Subchapter B of the enacted text (Sections 552.052 through 552.057). These provisions bind any person developing, deploying, or distributing AI in Texas or affecting Texas residents (not just government agencies), so this is where private-sector exposure actually lives.
Incitement to Self-Harm, Harm to Others, or Crime
Section 552.052 prohibits developing or deploying an AI system that "intentionally aims to incite or encourage" a person to commit physical self-harm, harm another person, or engage in criminal activity. The intent standard matters: this targets AI purpose-built or knowingly configured for incitement, not a general-purpose chatbot that a bad actor separately misuses.
Social Scoring
Section 552.053 prohibits AI that evaluates or classifies a person based on social behavior or predicted characteristics in a way that results in detrimental or unfavorable treatment, or that infringes constitutional rights, though this provision applies specifically to governmental entities, not private companies. A private-sector credit-scoring or insurance-underwriting model is governed by existing sector-specific law, not by TRAIGA's social-scoring clause.
Impairment of Constitutional Rights
Section 552.055 prohibits developing or deploying AI "with the sole intent" to infringe, restrict, or otherwise impair a person's constitutional rights. The "sole intent" language sets a high bar deliberately: it is not a general due-process or free-speech compliance obligation for ordinary AI products.
Unlawful Discrimination
Section 552.056 prohibits AI developed or deployed "with the intent to unlawfully discriminate against" a protected class under existing Texas or federal civil rights law. Critically, the statute states that "disparate impact is not sufficient by itself" to establish a violation — Texas requires proof of intent, which is a meaningfully lower private-sector exposure than a disparate-impact standard would create.
CSAM and Unlawful Deepfakes
Section 552.057 prohibits AI used to produce child sexual abuse material, or "deep fake videos or images" that violate Section 21.165 of the Texas Penal Code, as well as AI conducting text-based conversations simulating sexual conduct while impersonating a minor. This provision has the clearest, most direct criminal-law backing of the five and the least ambiguity in application.
Government Agency Disclosure Duties
Section 552.051(b) requires that a governmental agency offering a consumer-facing AI system disclose, before or at the time of interaction, that the consumer is interacting with an AI system. The disclosure must be "clear and conspicuous" and written in plain language, without dark patterns, though it may be delivered via hyperlink. Healthcare providers get a narrow carve-out: disclosure must happen no later than when treatment is first provided, except in emergencies.
This duty binds government agencies specifically. A private company operating a customer-facing AI chatbot is not captured by Section 552.051, a distinction that gets flattened in a lot of secondary coverage, and one worth confirming directly against your own deployment before assuming you're in scope.
Enforcement Mechanics: What Actually Happens If the AG Comes Calling
TRAIGA's enforcement design is deliberately narrow at the front door and serious at the back. Under Section 552.101, the Texas Attorney General has exclusive authority to enforce the chapter, with a limited exception for state agency licensing actions. There is no private right of action anywhere in the statute. An individual harmed by a violation cannot sue the developer or deployer directly under TRAIGA. Their recourse is the AG's consumer AI-rights complaint portal, which has been live and accepting public complaints since the law's January 1, 2026 effective date, and which functions as the practical trigger for an AG investigation. Filed complaints become public records under Texas open-records law, so a complaint against your organization is discoverable even before any enforcement action begins.
Section 552.103 gives the AG the power to issue civil investigative demands compelling a target to produce AI system descriptions, training data, performance metrics, known limitations, and post-deployment safeguards. Before any penalty can attach, Section 552.104 requires the AG to give written notice of the alleged violation, and the agency "may not bring an action" until at least 60 days after that notice, and not at all if the violation is cured and documented within that window. Capriglione confirmed this was a deliberate design choice: "We allow for an opportunity to cure the unlawful activity."
If a violation isn't cured, Section 552.105(a) sets three penalty tiers: curable violations run $10,000 to $12,000; uncurable violations run $80,000 to $200,000; and continuing violations run $2,000 to $40,000 per day. As of mid-2026, no formal TRAIGA enforcement action has been publicly filed, but the complaint portal is operating and the AG's investigative authority is fully active. The absence of a filed case is not the same as the absence of enforcement risk, particularly given how quickly a 60-day cure clock starts once a complaint prompts an investigation.
The statute also builds in a genuine affirmative defense that most compliance guides skip past. Section 552.105(e) says a defendant is not liable if the violation traces to a third party's unauthorized misuse of its system, or if the defendant discovered the issue through developer or deployer feedback, adversarial or red-team testing, adherence to applicable state agency guidelines, or "substantially complies with the most recent version of the 'Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile' published by the National Institute of Standards and Technology or another nationally or internationally recognized risk management framework." That last clause is the practical payoff: documented alignment with NIST AI RMF, or an equivalent recognized framework, isn't required by TRAIGA, but it is the single strongest defense available if the AG ever investigates, which is reason enough to keep that documentation current even though nothing in the statute mandates producing it up front.
The Texas AI Council and the Regulatory Sandbox
Chapter 554 creates the Texas Artificial Intelligence Council: seven voting members (three appointed by the governor, two by the lieutenant governor, two by the speaker of the House), serving staggered four-year terms, drawn from AI, data privacy, ethics, public policy, or risk-management backgrounds. The Council is administratively attached to the Texas Department of Information Resources. Its powers are advisory, not regulatory — it can issue reports, run agency training, and evaluate regulatory impediments, but it cannot adopt binding rules or interfere with agency operations.
Chapter 553 establishes a regulatory sandbox, run by the Department of Information Resources in consultation with the Council, allowing a company to test an innovative AI system for up to 36 months (extendable for good cause) without obtaining certain state licenses that would otherwise apply. Applicants must submit a system description, a benefit assessment addressing consumer, privacy, and safety impact, a mitigation plan, and proof of federal AI compliance, and must file quarterly reports once admitted. One limit matters more than the others: the sandbox waives licensing requirements, but the Subchapter B prohibited-use provisions stay fully binding even for sandbox participants. There is no path to legally test a system that violates the five prohibited uses.
For a genuinely novel AI deployment, the sandbox is a real compliance-friendly option — one worth building into your AI governance program as a documented decision point rather than an afterthought, since the application itself requires the same kind of risk and mitigation documentation your governance platform should already be producing.
A Documentation Checklist Scoped to What TRAIGA Actually Requires
Build your Texas-specific file around what the statute actually asks for, not a generic AI-governance template:
- Confirm whether your organization is a governmental entity or contracts as one for a Texas agency. If not, the Section 552.051 disclosure duty doesn't apply to you directly, though agency clients may push it into your contract.
- Map every AI system against the five Subchapter B prohibited uses specifically, not against a generic "high-risk" taxonomy. This is a narrower and faster exercise than a full EU AI Act risk classification.
- For any system with plausible discrimination exposure, document the intent standard: what your system is designed to do, and what safeguards prevent it from being used with unlawful discriminatory intent. Disparate-impact analysis alone doesn't establish or disprove a TRAIGA violation, so don't substitute one for the other.
- Keep a maintained record of AI system descriptions, training data provenance, and known limitations (the same categories a CID under Section 552.103 would demand) so you're not producing this under a 60-day cure clock for the first time.
- Document your alignment with the NIST AI Risk Management Framework's Generative AI Profile (or an equivalent recognized framework), even though TRAIGA doesn't require it up front. It's the strongest affirmative defense the statute offers under Section 552.105(e) if the AG ever investigates.
- If you're piloting a genuinely new AI capability that might brush against an existing Texas licensing requirement, evaluate the sandbox application before launch, not after a complaint.
- Track your multi-state posture against the actual 2026 enacted text of each law, not the ambitious draft version: Texas, Colorado, and California all narrowed materially between first draft and current effective law.
A governance maturity assessment that scores your program against the requirements you actually have (rather than a generic AI framework checklist) is the fastest way to confirm you're neither over-built nor exposed heading into any AG inquiry.
Common Issues and Fixes
"We assumed TRAIGA required private-sector risk assessments like Colorado's original bill." It doesn't, and hasn't since before enactment. Redirect the budget you'd allocated to a TRAIGA-specific assessment program toward the actual Subchapter B mapping exercise above. It's a smaller, faster task.
"We can't tell if our chatbot triggers the government disclosure duty." Check who operates the system, not who it serves. Section 552.051 binds governmental agencies; a private vendor building a chatbot for a state agency client should confirm contractually who carries that disclosure obligation, since the agency's duty can flow into vendor requirements even though the statute itself doesn't bind you directly.
"Legal wants to know if disparate-impact data alone creates exposure." It doesn't under Section 552.056, intent is required, and the statute says so explicitly. Don't let a compliance program built for a disparate-impact standard (like some EU or Colorado-style frameworks) get applied wholesale to your Texas obligations.
"We got a complaint filed against us on the AG portal and don't know what happens next." Nothing is automatic. The AG must independently investigate and, if it finds a violation, issue written notice starting a 60-day cure clock before any penalty can attach. Use that window: document the fix and submit supporting evidence before day 60, and the penalty doesn't attach at all.
Frequently Asked Questions
Does TRAIGA require private companies to conduct AI risk assessments?
No. The enacted version of TRAIGA dropped the broad private-sector risk-assessment regime that appeared in earlier drafts. Private-sector obligations are limited to not engaging in the five prohibited uses under Subchapter B; there is no general assessment, documentation, or audit mandate comparable to the EU AI Act or Colorado's original framework.
Who enforces TRAIGA, and can I be sued directly by someone I harmed?
The Texas Attorney General has exclusive enforcement authority. TRAIGA does not create a private right of action, so an individual cannot sue you directly under the statute. Their only avenue is filing a complaint through the AG's consumer AI-rights portal, which may (but does not automatically) trigger an AG investigation.
What is the 60-day cure notice, and does every violation get one?
Before the AG can bring an enforcement action, it must give written notice of the alleged violation, and it cannot file the action for at least 60 days. If the violation is cured and documented within that window, no penalty attaches. This applies to curable violations; the statute does distinguish curable from uncurable violation categories for penalty purposes.
How much can a TRAIGA violation actually cost?
Penalties range from $10,000 to $12,000 per curable violation, $80,000 to $200,000 per uncurable violation, and $2,000 to $40,000 per day for continuing violations, all set by Section 552.105(a) of the statute.
Does aligning with the NIST AI Risk Management Framework actually protect us under TRAIGA?
It can. Section 552.105(e) creates an affirmative defense if you substantially comply with the current NIST AI Risk Management Framework's Generative AI Profile, or an equivalent recognized framework, in addition to defenses for third-party misuse and violations caught through your own red-team or adversarial testing. This is a defense, not a mandate: TRAIGA doesn't require the assessment up front, but having it on file is the strongest protection the statute offers if the AG investigates.
Does TRAIGA apply to us if we're based outside Texas?
Yes, if your AI system is developed, deployed, or distributed in a way that reaches Texas residents or operations within the state. TRAIGA is not limited to Texas-headquartered organizations.
How is TRAIGA different from Colorado's AI Act?
Colorado's original 2024 law was a broad risk-based framework with duty-of-care and impact-assessment obligations for high-risk automated decision systems. After two delays, Colorado's May 2026 rewrite (SB 189) stripped out much of that structure in favor of narrower disclosure and transparency rules, effective January 1, 2027. TRAIGA never had that broad regime in its enacted form — it went straight to a prohibited-use model effective January 1, 2026.
Is the Texas AI Council a regulator?
No. The Council is an advisory body attached to the Department of Information Resources. It can issue reports, provide agency training, and recommend policy, but it cannot adopt binding rules or interfere with agency operations.
What is the regulatory sandbox, and does it excuse us from the prohibited-use rules?
The sandbox lets a company test a novel AI system for up to 36 months without certain state licensing requirements, administered by the Department of Information Resources with Texas AI Council input. It does not waive the five Subchapter B prohibited uses, which remain fully enforceable even for sandbox participants.
If your AI governance program is still calibrated to the broadest early drafts of state AI law rather than what's actually in force, that gap is exactly what an AI Governance module built to track regulation-by-regulation is for — mapping each system against Texas's actual five prohibited uses, not a generic risk taxonomy, and keeping the documentation an AG civil investigative demand would ask for already on file before you're inside a 60-day clock.
